Next Guest Engine : SP3-NNUE

[chrisw]

Hello to the veteran of chess computer programming—I did not express any criticism; rather, I was asking for an explanation.
Furthermore, I made a request to make these output messages toggleable. There may well be some individuals for whom this output provides useful information; however, for the majority of users, it is pointless. Hence my request to be able to switch it off—something which, I assume, would be very simple to implement.

Then you would make posts claiming the program had crashed.

[Uri Blass]

SP3-NNUE is our third guest engine. It's made from Senpai 3.02 from Fabien Letouzey using self play games mixed with some Leela data using a new NNUE technique as an experiment. A tribute to Fabien's overall great work. SP3-NNUE is rated 3450-3500 elo. A quick match revealed :

Code: Select all

No. Name         Win Draw Loss Unf.  Score Games       %
--------------------------------------------------------
  1 SP3-NNUE     +93  =10   -0   *0   98.0   103   95.1%
  2 Senpai-3.02   +0  =10  -93   *0    5.0   103    4.9%

While Guest engines use the same search of Rebel and Cstal the NNUE evaluation should be unique and not be above 60% see our similarity report.

https://rebel7775.wixsite.com/rebel/guest-engines

1)I see the small similiarity between the different NNUE engines(less than 60%).
I would like to see also what is the similiarity between SP3-NNUE and Senpai-3.02 and same for similiar pares of HCE and NNUE engine

2)When I am not surprised by the fact that the NNUE version won I am surprised to see result near 95%

The result of the match between the engine based on shredder13's games and Shredder13 was only 85.2% and it is not that shredder13 is so much stronger relative to Senpai

From CCRL:

Shredder 13 64-bit 3246 +6 −6 46.8% +20.8 56.9% 6709
Senpai 3.0.2 64-bit 3222 +15 −15 48.1% +12.2 54.7% 953

[peter]

I'd support most of these questions too, even more I'd like to know, how comes that SP3 NNUE wins clearly against ChessSystemTal Extreme?
Had a match running single thread with UHO 2024 85-94 cp 6mvs 3'+1" and after about 150 decided games (in the 186 counted there are 30 unfinished ones because of stopping the match manually) I get this:

Score of CSTE vs SP3NNUE: 14 - 114 - 28 [0.179]
Elo difference: -264.0 +/- 60.5, LOS: 0.0 %, DrawRatio: 17.9 %
186 of 500 games finished.

Didn't have a closer look into the games so far, there could have been well some flaw with the CSTE- installation, even if the few games I watched, didn't show such as for obvious things like crashes or time- forfeits. If I don't see such systematic reasons on my own backwarding some of the games, I' could upload all of them in case of demand, regards

[Peter Berger]

Have we +really+ missed this kind of chess ? It definitely feels fruity.

Unique it is of course, it doesn't happen often to see four queens on the board that don't vanish at once. This game must have been a draw at a relatively late point in the game, but without just letting the engines figure it out, I can't pinpoint it at all-

15. ..Rde8 looked pretty awkward to me, 22. ...a5 came as a huge personal surprise, difficult to understand at first.

Given that these engines share quite a lot, it is just amazing how very different they feel and behave style-wise. There was a time when we had to pay like 100€ for this kind of new experience, I really like the new ones.

[pgn]

[Event "Lang 120min+10sek"]
[Site "Berlin"]
[Date "2026.05.09"]
[Round "?"]
[White "Crafty 25.6"]
[Black "SP3-NNUE"]
[Result "0-1"]
[ECO "B85"]
[PlyCount "228"]
[TimeControl "7200+10"]

{4096MB, DESKTOP-8OCGGEO} 1. e4 {[%emt 0:00:00]} c5 {[%eval 32,31] [%emt 0:02:46]} 2. Nf3 {[%emt 0:00:06]} e6 {[%eval 60,27] [%emt 0:01:27]} 3. d4 {[%emt 0:00:07]} cxd4 {[%eval 48,26] [%emt 0:00:53]} 4. Nxd4 {[%emt 0:00:07]} Nc6 {[%eval 51,29] [%emt 0:03:55]} 5. Nc3 {[%emt 0:00:06]} Qc7 {[%eval 77,26] [%emt 0:01:16]} 6. Be2 {[%emt 0:00:07] (Le3)} Nf6 {[%eval 36,26] [%emt 0:00:54]} 7. O-O {[%emt 0:00:07] (Sdb5)} Be7 {[%eval 29,28] [%emt 0:00:51]} 8. Be3 {[%emt 0:00:07]} O-O {[%eval 37,31] [%emt 0:00:58]} 9. f4 {[%emt 0:00:07]} d6 {[%eval 29,28] [%emt 0:02:18]} 10. Kh1 {[%emt 0:00:07]} a6 {[%eval 34,26] [%emt 0:00:52]} 11. Qe1 {[%emt 0:00:08]} Nxd4 {[%eval 0,25] [%emt 0:00:56]} 12. Bxd4 {[%emt 0:00:06]} b5 {[%eval 39,25] [%emt 0:03:08]} 13. a3 {[%emt 0:00:06] (Dg3)} Bb7 {[%eval -31,25] [%emt 0:02:24]} 14. Qg3 {[%emt 0:00:08] (Lf3)} Rad8 {[%eval 0,26] [%emt 0:04:57]} 15. Rae1 {[%emt 0:00:10] (h3)} Rde8 {[%eval 26,27] [%emt 0:03:08]} 16. Rd1 {[%emt 0:08:09] (Lf3)} h6 {[%eval 19,27] [%emt 0:03:51]} 17. Rd3 {[%emt 0:03:12] (Tfe1)} Rd8 {[%eval 13,29] [%emt 0:02:45]} 18. Bf3 {[%emt 0:00:13] (Te1)} Rd7 {[%eval 31,29] [%emt 0:02:31]} 19. Rfd1 {[%emt 0:02:53] (Te1)} Rfd8 {[%eval 26,28] [%emt 0:02:05]} 20. Qf2 {[%emt 0:00:29] (Te1)} Re8 {[%eval 10,32] [%emt 0:15:20]} 21. Bb6 {[%emt 0:03:21]} Qc8 {[%eval 20,31] [%emt 0:01:41]} 22. Qg3 {[%emt 0:08:29]} a5 {[%eval 12,32] [%emt 0:06:23]} 23. e5 {[%emt 0:02:29] (Lxa5)} dxe5 {[%eval -10,30] [%emt 0:01:33]} 24. fxe5 {[%emt 0:00:36]} Rxd3 {[%eval 0,36] [%emt 0:01:47]} 25. Rxd3 {[%emt 0:01:26] (cxd3)} Nd7 {[%eval -75,29] [%emt 0:01:17]} 26. Bxa5 {[%emt 0:02:44]} Bxf3 {[%eval -23,32] [%emt 0:02:25]} 27. gxf3 {[%emt 0:02:09] (Txf3)} Nf8 {[%eval -58,28] [%emt 0:01:29]} 28. h4 {[%emt 0:02:01] (Df2)} h5 {[%eval -93,29] [%emt 0:02:12]} 29. Qh2 {[%emt 0:03:54]} Ng6 {[%eval -76,31] [%emt 0:07:41]} 30. Ne4 {[%emt 0:01:58]} Qa6 {[%eval -84,33] [%emt 0:02:42]} 31. Bc7 {[%emt 0:01:56] (Le1)} Rc8 {[%eval -81,32] [%emt 0:01:36]} 32. Bd6 {[%emt 0:01:48]} Bxh4 {[%eval -76,32] [%emt 0:00:36]} 33. Bc5 {[%emt 0:01:02] (b3) R} Qc6 {[%eval -138,29] [%emt 0:02:39]} 34. c3 {[%emt 0:01:56]} Qa8 {[%eval -102,28] [%emt 0:00:56]} 35. Bd6 {[%emt 0:00:36]} Rc4 {[%eval -106,28] [%emt 0:01:29]} 36. Rd1 {[%emt 0:00:07]} Qa4 {[%eval -117,28] [%emt 0:01:09]} 37. Rf1 {[%emt 0:03:32] (Dd2)} Bd8 {[%eval -126,25] [%emt 0:00:48]} 38. Re1 {[%emt 0:03:01] (De2)} Nh4 {[%eval -132,29] [%emt 0:00:49]} 39. Qe2 {[%emt 0:01:38]} Bb6 {[%eval -114,30] [%emt 0:01:12]} 40. Bb4 {[%emt 0:01:28] (Kh2)} Nf5 {[%eval -163,28] [%emt 0:01:00]} 41. Qd2 {[%emt 0:03:07] (Dd3)} Qa8 {[%eval -166,27] [%emt 0:00:48]} 42. Qd3 {[%emt 0:01:11] (Kg2)} Qb8 {[%eval -215,29] [%emt 0:00:56]} 43. b3 {[%emt 0:00:37]} Rxe4 {[%eval -172,33] [%emt 0:01:05]} 44. Rxe4 {[%emt 0:00:08]} Ng3+ {[%eval -183,32] [%emt 0:00:54]} 45. Kh2 {[%emt 0:00:13]} Nxe4 {[%eval -193,33] [%emt 0:00:44]} 46. Qxe4 {[%emt 0:01:20]} Qd8 {[%eval -177,34] [%emt 0:01:32]} 47. c4 {[%emt 0:00:08]} Qd1 {[%eval -148,30] [%emt 0:00:46]} 48. c5 {[%emt 0:00:20]} Bc7 {[%eval -138,31] [%emt 0:01:17]} 49. Qa8+ {[%emt 0:00:08] (De3)} Kh7 {[%eval -205,29] [%emt 0:00:39]} 50. Qe4+ {[%emt 0:05:22]} Kg8 {[%eval 0,37] [%emt 0:01:18]} 51. Qa8+ {[%emt 0:01:05]} Kh7 {[%eval 0,4] [%emt 0:00:01]} 52. Qe4+ {[%emt 0:01:09]} g6 {[%eval -185,37] [%emt 0:00:34]} 53. Qe3 {[%emt 0:00:18]} Qc2+ {[%eval -205,34] [%emt 0:00:33]} 54. Bd2 {[%emt 0:01:56] (Kg1)} Qb2 {[%eval -224,30] [%emt 0:00:36]} 55. f4 {[%emt 0:00:13]} Qxa3 {[%eval -205,32] [%emt 0:00:46]} 56. Kg3 {[%emt 0:00:07] (Le1)} Qa1 {[%eval -239,30] [%emt 0:00:53]} 57. b4 {[%emt 0:00:06] (Le1)} Qh1 {[%eval -240,33] [%emt 0:00:30]} 58. Qf3 {[%emt 0:00:19]} Qg1+ {[%eval -240,35] [%emt 0:00:58]} 59. Kh3 {[%emt 0:03:09]} Kg7 {[%eval -240,40] [%emt 0:01:04]} 60. Bc3 {[%emt 0:00:52]} Qb1 {[%eval -240,36] [%emt 0:00:39]} 61. Kg2 {[%emt 0:01:22] (Kh2)} Qc1 {[%eval -240,37] [%emt 0:00:55]} 62. Qg3 {[%emt 0:01:26] (Kh2)} Qc2+ {[%eval -282,31] [%emt 0:00:42]} 63. Kf1 {[%emt 0:00:12]} Qe4 {[%eval -278,32] [%emt 0:00:13]} 64. Be1 {[%emt 0:01:28] (Ld2)} Bd8 {[%eval -280,35] [%emt 0:00:44]} 65. Bd2 {[%emt 0:00:49]} Qh1+ {[%eval -275,36] [%emt 0:02:02]} 66. Ke2 {[%emt 0:00:43]} Qd5 {[%eval -275,35] [%emt 0:02:59]} 67. Qh3 {[%emt 0:01:19]} Kf8 {[%eval -275,35] [%emt 0:00:30]} 68. Ke1 {[%emt 0:00:07] (Dg3)} Kg8 {[%eval -275,33] [%emt 0:00:19]} 69. Qe3 {[%emt 0:02:56] (Ke2)} Qh1+ {[%eval -278,35] [%emt 0:00:29]} 70. Ke2 {[%emt 0:00:33]} Qg2+ {[%eval -337,35] [%emt 0:00:33]} 71. Ke1 {[%emt 0:00:08] (Df2)} Qh1+ {[%eval 0,5] [%emt 0:00:01]} 72. Ke2 {[%emt 0:00:33]} Qg2+ {[%eval 0,4] [%emt 0:00:01]} 73. Ke1 {[%emt 0:00:28]} Bh4+ {[%eval -337,31] [%emt 0:00:39]} 74. Kd1 {[%emt 0:00:07]} Be7 {[%eval -345,33] [%emt 0:00:19]} 75. Ke1 {[%emt 0:00:06] (De2)} Bh4+ {[%eval 0,8] [%emt 0:00:00]} 76. Kd1 {[%emt 0:01:06]} Be7 {[%eval 0,4] [%emt 0:00:01]} 77. Ke1 {[%emt 0:00:23]} Kf8 {[%eval -349,34] [%emt 0:00:28]} 78. Qf2 {[%emt 0:00:35]} Qh3 {[%eval -349,33] [%emt 0:00:26]} 79. Ke2 {[%emt 0:00:38] (Df1)} Qg4+ {[%eval -374,32] [%emt 0:00:21]} 80. Kd3 {[%emt 0:01:13] (Df3)} h4 {[%eval -473,25] [%emt 0:00:25]} 81. Kc3 {[%emt 0:02:03] (c6)} h3 {[%eval -604,27] [%emt 0:00:29]} 82. c6 {[%emt 0:00:28]} Bd8 {[%eval -631,30] [%emt 0:00:21]} 83. Be3 {[%emt 0:00:18]} Kg7 {[%eval -660,30] [%emt 0:00:26]} 84. Bc5 {[%emt 0:00:30]} Bh4 {[%eval -657,30] [%emt 0:00:20]} 85. Qf1 {[%emt 0:00:21]} Qg3+ {[%eval -691,29] [%emt 0:00:14]} 86. Qd3 {[%emt 0:00:07]} Qxf4 {[%eval -691,27] [%emt 0:00:02]} 87. c7 {[%emt 0:00:11]} h2 {[%eval -711,30] [%emt 0:00:30]} 88. c8=Q {[%emt 0:00:06]} h1=Q {[%eval -707,29] [%emt 0:00:11]} 89. Kb3 {[%emt 0:00:08]} Qa1 {[%eval -726,30] [%emt 0:00:18]} 90. Bf8+ {[%emt 0:00:23]} Kh7 {[%eval -779,33] [%emt 0:00:49]} 91. Qcc3 {[%emt 0:00:42] (Dcc2)} Qxc3+ {[%eval -762,26] [%emt 0:00:10]} 92. Qxc3 {[%emt 0:00:27]} Bg5 {[%eval -809,32] [%emt 0:00:11]} 93. Qd3 {[%emt 0:01:30] (Ld6)} Qxe5 {[%eval -1154,27] [%emt 0:00:17]} 94. Kc2 {[%emt 0:00:07] (Ld6)} Qf5 {[%eval -1256,29] [%emt 0:00:16]} 95. Bd6 {[%emt 0:00:18] (Dxf5)} Bf6 {[%eval -1402,30] [%emt 0:00:15]} 96. Bc5 {[%emt 0:00:18] (Dxf5)} g5 {[%eval -1552,30] [%emt 0:00:30]} 97. Bd6 {[%emt 0:00:42] (Le3)} Kg6 {[%eval -1722,31] [%emt 0:00:18]} 98. Qxf5+ {[%emt 0:00:30] (Kd2)} Kxf5 {[%eval -1847,30] [%emt 0:00:14]} 99. Kb1 {[%emt 0:00:28] (Kd3)} Be5 {[%eval -2809,26] [%emt 0:00:15]} 100. Bc5 {[%emt 0:00:26] (Le7)} Ke4 {[%eval -32737,60] [%emt 0:00:10]} 101. Kc2 {[%emt 0:00:11] (Kc1)} f5 {[%eval -32739,59] [%emt 0:00:05]} 102. Bb6 {[%emt 0:00:10]} g4 {[%eval -32741,64] [%emt 0:00:31]} 103. Kd2 {[%emt 0:00:06] (Kc1)} f4 {[%eval -32743,76] [%emt 0:00:05]} 104. Kd1 {[%emt 0:00:05] (Ke1)} g3 {[%eval -32745,73] [%emt 0:00:05]} 105. Kd2 {[%emt 0:00:05] (Ke1)} f3 {[%eval -32747,82] [%emt 0:00:05]} 106. Ke1 {[%emt 0:00:05]} Bd4 {[%eval -32749,107] [%emt 0:00:01]} 107. Bxd4 {[%emt 0:00:06]} Kxd4 {[%eval -32751,116] [%emt 0:00:01]} 108. Kf1 {[%emt 0:00:06]} Ke3 {[%eval -32753,116] [%emt 0:00:00]} 109. Ke1 {[%emt 0:00:09]} g2 {[%eval -32755,116] [%emt 0:00:01]} 110. Kd1 {[%emt 0:00:05]} Kd3 {[%eval -32757,116] [%emt 0:00:00]} 111. Kc1 {[%emt 0:00:06]} Kc3 {[%eval -32759,116] [%emt 0:00:00]} 112. Kb1 {[%emt 0:00:07]} g1=Q+ {[%eval -32761,116] [%emt 0:00:01]} 113. Ka2 {[%emt 0:00:07]} Qf2+ {[%eval -32763,116] [%emt 0:00:01]} 114. Ka3 {[%emt 0:00:07] (Ka1)} Qa7# {[%eval -32765,116] [%emt 0:00:01]} 0-1

[/pgn]

[Rebel]

I'd support most of these questions too, even more I'd like to know, how comes that SP3 NNUE wins clearly against ChessSystemTal Extreme?
Had a match running single thread with UHO 2024 85-94 cp 6mvs 3'+1" and after about 150 decided games (in the 186 counted there are 30 unfinished ones because of stopping the match manually) I get this:

Score of CSTE vs SP3NNUE: 14 - 114 - 28 [0.179]
Elo difference: -264.0 +/- 60.5, LOS: 0.0 %, DrawRatio: 17.9 %
186 of 500 games finished.

Didn't have a closer look into the games so far, there could have been well some flaw with the CSTE- installation, even if the few games I watched, didn't show such as for obvious things like crashes or time- forfeits. If I don't see such systematic reasons on my own backwarding some of the games, I' could upload all of them in case of demand, regards

Because you were not playing with the strongest CSTAL but with the EXTREME version which is not so much about elo but all about playing style.